Respiration Estimation and Apnea Detection Using Fuzzy Logic
Akrit Mudvari, Student Member, IEEE and Taikang Ning, Senior Member, IEEE
Abstract— This paper presents a fuzzy logic-based
respiration monitoring algorithm that is capable of providing
accurate respiration rate and detecting apnea episodes. The
proposed algorithm employs several signal processing
techniques to extract useful features that signify different
respiratory behaviors. We implement a fuzzy logic-based
system that examines the extracted respiratory signal features
and categorizes the respiratory signals into respiration, body
motion, and apnea. The performance of the underlying
algorithm is validated using both the MIT physiology database
and in-house respiration measurements.
I. INTRODUCTION
Sleep apnea syndrome is a serious health threat problem
that affects 5% of the world’s total population. The
percentage is even higher among certain demographics. For
example, it rises up to 30% for males over the age of 70 [1].
Apnea is known to result in multiple problems including
abnormal behavior during sleep such as body movements
resulting in injuries, nocturnal enuresis, headache in the
morning, and drowsiness during the day due to lack of proper
sleep [2]. While sleep apnea is a prevalent disease, clinical
diagnosis of sleep apnea still has a lot of room for
improvement [3].
In this paper, four signal processing methods were used in
our study to extract multiple respiration signal features.
Adaptive approach was utilized to dynamically adjust the
thresholding values to gain better accuracy in respiration rate
estimation. For instance, adaptive threshold was used to
account for loosening of the abdominal strain gauge, which is
an impedance measurement device strapped to an elastic
band [6]. This is a necessary measure because respiration rate
varies by the health condition [6] and the patient’s age.
Once the breathing activity is recognized using these
methods, respiration rates measured using different methods
and signal strength are used to detect the definite instances of
apnea. Here, higher preference is allocated to eliminating
false negatives because a real time detection would enable
immediate medical attention. However, it is necessary to
realize that with the onset of apnea, body struggles could
result in some activity in strain gauge, which resembles but is
not identical to body movements during normal breath [8].
Since it is more difficult to separate normal breathing activity
from signals effected by the noises due to body movement,
fuzzy logic-based approach is employed to process the signal
and classify a period of respiration into normal breathing and
body movement. Episodes with overlapping time periods are
taken to minimize the chances of missing out on instances of
apnea, and data obtained for different source are used to
verify that the apnea is detected in each case.
_____________________________
A. Mudvari and T. Ning are with the Engineering Department of Trinity
College, Hartford, CT 06106, USA (tel: 860-297-2517; fax: 860-297-2531;
e-mail: akrit.mudvari@trincoll.edu)
978-1-5090-2809-2/17/$31.00 ©2017 IEEE
II. METHOD
To provide an accurate estimation of the respiration rate
and to correctly detect the apnea episode, we have employed
several signal processing methods to extract key features that
score the respiration behavior.
A. Adaptive Thresholding
In this method, breath information is extracted based on
the waveform fluctuations across an adaptive threshold. This
threshold is dynamically adjusted using most recently
detected breathing features. Weighted average of previous
breaths is calculated, where most recent breaths are assigned
more weight. The main advantage of an adaptive threshold is
that the breath waveform changes due to body movement or
loosening of respiration transducer during sleep will have
much less effect on accuracy of the respiration rate
estimation. The threshold of adaptive thresholding is
initialized and calibrated using several normal breaths from
the human subject.
In case of respiration with body movement, the signal
resembles amplitude variation of normal breathing but
exhibits aperiodic and irregular waveform. Time period
between consecutive signals of breath is taken into account
and anomalous data is excluded. As such, the processed
result enables us to know whether the recorded signal is
caused by normal respiration or motion artifacts. Fig.1
exemplifies the classification of detected peaks into normal
breathing and breathing with body motion artifacts.
B. Zero-crossing and Signal Energy
In the second process, strength of the signal is acquired in
real time, and a few seconds-long samples are processed to
have zero average. In order to acquire a zero-average,
weighted averages of the most recent data are taken and the
weights are assigned to allocate more importance to the most
Figure 1: Adaptive thresholding detection: detected breaths are
noted by as vertical lines where lines having amplitude near 200
represent body movement noise.
2818
recent data. This ensures that loosening or tightening of the
strain gauge does not reduce accuracy of the algorithm. Zero
crossing method allows the algorithm to detect and
distinguish “normal breath” and “breath with movement” in
real-time and calculate the respiration rate. A difference
between this technique and the threshold crossing method is
that here, the standard deviation of the data is used to
measure the threshold, rather than the average.
Another used of the zero averaged data is calculation of
the signals “energy” (root mean squared of the signal’s zero-
averaged amplitude for a particular sample collected in real
time). This method is particularly useful in detecting apnea
because the absence of adequate expansion and contraction of
abdominal strain gauge means that there is neither normal
breathing nor any signal generation due to body movements.
Figure 2: Membership function of the relative magnitude of
C. Respiration Waveform Slope dominant frequency for normal respiration (solid line) and body
movement artifacts (dashed line)
This method uses the magnitude and the rate of change in
respiration amplitude to categorize the signal into normal A. Membership Function Using AR Coefficient Magnitude
respiration, apnea, and body movement artifacts. Normal
After obtaining data on episodes that are definitively
breathing signal is used to obtain the baseline data that
normal (based on the observation of healthy breathing
characterize normal respiration. The “rise” and “fall” of a
episodes), the average and the standard deviation of relative
breath signal are noted by their associated “slopes.” These
magnitude of the dominant frequency are obtained. Same
waveform slope features are included in the proposed
process is repeated for episodes that were classified as a
algorithm to classify respiration waveform signals.
result of body movement. The membership functions as
D. Autoregressive (AR) modelling shown in Fig.2 are constructed using the obtained values.
Input membership functions (IM) for normal breath is
The respiration rate can be efficiently estimated using obtained using the following formula:
second order autoregressive (AR) model. While higher order
AR models are also applicable, it is sufficient to use a
second-order AR model (1) to capture the dominant
frequency of a short data segment [9]. (3)

e k  x ( k )  a1 x ( k  1)  a 2 x ( k  2 ) (1)
Similarly, IMmovement is calculated with 1 for x<µmovement.
AR model coefficients {a1, a2} are estimated using a short
data segment and the respiration rate is estimated by B. Membership Function Using Peak Average of ‘normal’
and ‘movement artifact’:
2
f samp 4 a 2  a1 Once the respiration rates for signals classified as ‘normal’
). (2)
1
freq  tan ( and as ‘affected with movement’ are obtained, membership
2 a1
functions are heuristically assigned (the assignment emulated
where fsamp is the sampling frequency used during respiration decision-making of the experts and was later statistically
measurement. verified). Table I. shows the example of assigned
Since the normal breath signals were found to exhibit membership function values.
regular behavior similar to a sinusoidal signal while body
movement artifacts displayed irregular fluctuations, the 2nd
order AR model refection coefficient (a2) can be used to TABLE I: ASSIGNED MEMBERSHIP FUNCTIONS OF NORMAL RESPIRATION AND
signify the closeness to a sine wave, where the magnitude of MOTION ARTIFACTS
a2 closer to 1 indicates a regular waveform.
Motion Peaks Normal Respiration Peaks
0 1 2 3 4
III. FUZZY LOGIC APPROACH
0 0 0.1 0.6 0.9 1
A Fuzzy-logic [9] based approach is used in our study to
1 0 0.1 0.5 0.85 0.96
analyze respiration signals. We developed membership
functions that assign varying degrees of membership to the 2 0 0.1 0.45 0.85 0.93
variables that are used in decision-making [9]. The first input
3 0 0.07 0.45 0.83 0.88
variable used in the calculation is the relative magnitude of
the dominant frequency obtained from autoregressive >=4 0 0.05 0.35 0.8 0.85
modelling. For the second input, average of healthy peaks
detected using adaptive threshold method, zero crossing
method and respiration waveform slope method are used.
2819
 Figure 3: Detection of apnea (apnea is signified by a lack of signal
strength). The program increases the amplitude (of dashed line in the
figure) to about 180 when apnea episode is detected.
C. Fuzzy Inference and Classification
Based on the aforementioned membership functions,
output memberships are calculated using fuzzy inference
engine [11]. An algorithm is used where the following
equation gives output membership function for normal breath
) rate (detected using threshold crossing method, zero-crossing
Where, OM refers to output membership function,
ARMAG refers to relative magnitude of the dominant
frequency and PD refers to detected peaks. a and b are the
weights assigned to each membership function (a+b=1). The
values of a and b were obtained after statistical analysis of the
acquired set of data, and these values can be further
strengthened with further study and use of more data. The
following equation is used to calculate the output
membership function (OMF) for movement:
(5)
Each output membership function is compared to decide
whether the episode should be categorized as normal breath
episode or episode with body movements.
iT
IV. RESULTS AND DISCUSSION
To examine the performance of the proposed algorithm,
respiration signals measured by an abdominal strain gauge
were used [4] with a sampling rate of 20 Hz. Respiratory
measurements were classified as ‘breathing with instances of
apnea’, ‘normal breathing’ and ‘breathing with body
movement’. The algorithm was designed to process
respiratory data in short segments of duration of 10 seconds
or 15 seconds. Each short segment will be categorized
according to the extracted respiration features.
Apnea is characterized by a lack of breathing for a period
of 10 seconds or longer; an average of 25 seconds was noted
in the literature [3]. In our algorithm, we used an interval of
20 seconds as a time frame for a respiration segment for
analyzing the data and classifying the episodes into apnea,
normal breath or body movements. The algorithm is also run
2820
Fig 4: Detection of episodes of body movement (noisy signal).
Detection of such episodes is shown by horizontal line with non-zero
amplitude (~50).
by overlapping 50% of the episode with adjacent segments.
For instance, the first 20 seconds are considered to compose
first episode while the second episode covers 10-30 seconds.
The 50% overlapping increases the detection of short apnea
episodes. Two key parameters were used to ensure that every
instance of apnea is detected. They are strength of the signals
(i.e., root mean square of energy) and the average respiration
method and waveform slope measurement method). If either
technique detects an instance of apnea, it would be safe to
assume that the analyzed episode represents apnea.
For data containing mostly respiration activities and
apnea episodes in between, our algorithm successfully
detected all apnea instances. The more complicated situations
were when either possible apnea episodes were distorted with
significant body movement artifacts, or when the data
segment under examination was the combination of apnea
and respiration. The former was categorized as “body
struggle” by the algorithm. When we changed the disjoint
20-second data segments into 50% overlapped data segments,
the precision of detecting the onset of apnea was improved.
Fig.3 shows that our algorithm can accurately detect the onset
of apnea.
It was observed from data that respiratory measurement
caused by body movements was characterized by a lack of
signal regularity, inconsistency in frequency and amplitude
recording, and a sudden change of the waveform. To classify
such waveforms, fuzzy logic-based decision making was
used. Fig. 4 shows a set of signals resulting from body
movements, and our algorithm was able to correctly
categorize these instances as body movements. In case of the
body movements, our fuzzy-logic based heuristic
classification algorithm achieved more than 90% agreement
with human experts. In case of episodes where the human
judgment and the fuzzy logic based algorithm differed, the
input membership functions for “normal breathing” and
“instances of body movement” were found to be very close to
each other. Here, expert opinion and the use of more data for
statistical treatment could make the decision making process
even more robust.

Figure 5. Detection of the onset of body movements (after
around 10 seconds). Detection of the episodes with body
movements is signified by horizontal dashed line with non-zero
amplitude.
As Fig. 5 illustrates, the proposed algorithm is capable of
separating episodes of healthy breathing from episodes with
significant disturbances of body movement. The results were
agreeable with human experts. In Fig. 6, we can see that the
algorithm is capable of detecting apnea and body movement.
The performance of our approach was also validated
using the PhysioBank MIT database [10]. We used both
healthy breathing and apnea episodes. The algorithm was
successful in detecting all apnea episodes as shown in Fig.7.
V. CONCLUSION
This paper demonstrates the performance of a new
respiration monitoring system. In absence of the apnea
episodes, our algorithm was able to accurately estimate the
respiration rate. Our results have shown that respiratory
signal features extracted from multiple signal processing
methods can be effectively analyzed in real time using a
fuzzy logic-based expert system to correctly categorize
respiratory signals into normal breathing, breathing with
motion artifacts, and apnea.
Figure 7. Detection of normal breathing (left) and detection of apnea (right) for signal collected from MIT online database [10]. Discussed
algorithm shifts the horizontal dashed line from amplitude of 0 to 1.5 when it detects apnea, as shown above.
2821
Figure 6. Real-time detection of i) body movement artifacts (shown by
low-amplitude (~50) horizontal dashed line for a duration of 15-35
seconds, and ii) apnea (shown by high amplitude (~200) horizontal
dashed line for a duration of 75-95 seconds).
REFERENCES
[1] P. Varady ; T. Micsik ; S. Benedek ; Z. Benyo, “A novel method for
the detection of apnea and hypopnea events in respiration signals.”,
IEEE Trans.Biomedical Eng., Volume: 49, Issue: 9, Sept. 2002.
[2] C. Guilleminault; A. Tilkian; W. C. Dement, “The Sleep Apnea
Syndromes”, Annual review of medicine, 1976, 27(1), 465-484.
[3] T. Young; L. Evans; L. Finn;M Palta, “Estimation of the clinically
diagnosed proportion of sleep apnea syndrome in middle-aged men
and women.” Sleep, 1997, 20, 705–706. Medline.
[4] T. Ning and J. D. Bronzino, "Autoregressive and bispectral analysis
techniques: EEG applications," Special Issue on Biomedical Signal
Processing for IEEE Engineering in Medicine and Biology Magazine,
Vol.9, No.1, pp.47-50, March 1990.
[5] K. Nepal ; E. Biegeleisen ; T. Ning, “Apnea detection and respiration
rate estimation through parametric modelling.”, Bioengineering
Conference, Proceedings of the IEEE 28th Annual Northeast, 2002
[6] Farah Q Al-Khalidi, Reza Saatchi, Derek Burke, Heather E Elphick,
Stephen Tan. Respiration Rate Monitoring Methods: A Review.
Pediatric Pulmonology, Wiley, 2011, 46 (6), pp.523.
[7] L.A. Wallis; M. Healy; M.B. Undy, I. Maconochie. “Age related
reference ranges for respiration rate and heart rate from 4 to 16 years.”
Arch Dis Child, 2005;90:1117-21.
[8] M. Nishyama; M. Miyamoto; K. Watanabe, “Respiration and body
movement analysis during sleep in bed using hetero-core fiber optic
pressure sensors without constraint to human activity.” J. Biomed.
Opt. 2011, Volume 16, Issue 1.
[9] L. A. Zahed, “Fuzzy Sets”, Information and Control, 1965, volume 8,
338-358
[10] A.L. Goldberger, L.A.N. Amaral, L Glass, et al.“PhysioBank,
PhysioToolkit, and PhysioNet: Components of a New Research
Resource for Complex Physiologic Signals.” Circulation 101(23),
2014, e215-e220